Automatic classification of papers related to carbon pricing using LLAMA2

Abstract

In order to investigate the analytical capability of LLAMA2 for environment-related datasets, we attempted to automatically classify the content of academic papers related to carbon pricing using LLAMA2. In the task of predicting whether the abstract of an academic paper is related to carbon pricing or not, we confirmed that LLAMA2 can stably generate output in the specified format and classify with a certain accuracy (F1 score: 0.66) that exceeds randomness, using only prompting. Compared to classification using BERT, the advantages of this method are that it can output results without fine tuning, thus reducing labor and cost, and that it can output the reason for classification in natural language. Another advantage over closed LLM classification by enterprise services such as ChatGPT is that it operates in a local environment, reducing the risk of information leakage, etc.